Re: Question about RTL for bitwise AND

[Ian Lance Taylor <ian@airs.com>]

"Matt Lee" <reachmatt.lee@gmail.com> writes:

> Here is what I have,
> 
>     case AND:
>     case IOR:
>     case XOR:
>         *total = COSTS_N_INSNS (1);
>         return true;
>     case ASHIFT:
>     case ASHIFTRT:
>     case LSHIFTRT:
>          if (GET_CODE (XEXP (x, 1)) == CONST_INT)
>              *total = COSTS_N_INSNS (INTVAL (XEXP (x, 1)));
>          return true;
> 
> 
> This looks quite OK to me. I tried debugging if rtx_costs was doing
> something wrong. I do see rtx_cost being invoked for the lshiftrt
> expressions in question, but never for the "and". Seems like GCC had
> pre-decided to only use lshiftrt even though it is expensive.

Your computations look wrong to me.  You are saying that an AND always
costs 1 instruction.  But you need to actually look at the operands.
Specifically, the compiler is going to compare these two:
    (and (reg ...) (const_int ...)
    (and (ashiftrt (reg ...) (const_int ...)) (const_int 1))

> Any other ideas? Btw, I couldn't find prefer_and_bit_test in dojump.c.

I'm looking at the current version of gcc.  prefer_and_bit_test was
added 2004-03-20, so I guess that just missed 3.4.  I don't have 3.4
sources handy.  In general the decision is made in the do_jump
function.

Ian